ClawMonitor

English | 简体中文

Keyboard-first OpenClaw monitor with:

• Per-session last inbound user message + last outbound assistant message (preview + timestamp)
• Per-model health probes (direct provider API and through OpenClaw itself)
• On-demand session history with task trajectory (todo / doing / done)
• Token visibility from local session snapshots plus Gateway 1d / 7d / 30d usage windows
• Gateway service / cgroup health view with zombie-orphan-helper inspection and reclaim estimates
• Work state: WORKING / FINISHED / INTERRUPTED / NO_MESSAGE (+ NO_FEEDBACK alert)
• Long-run visibility via *.jsonl.lock (works even if Gateway is down)
• Optional Gateway log tail + channel runtime snapshot correlation (Feishu/Telegram-focused rules)
• Full-screen TUI with manual “nudge” (send a progress request via chat.send)

Install (editable)

cd ~/program/clawmonitor
python3 -m pip install -e .

Install (PyPI)

pip install clawmonitor

Run

clawmonitor init
clawmonitor tui

Other commands:

clawmonitor snapshot --format json
clawmonitor snapshot --format md
clawmonitor nudge --session-key 'agent:main:main' --template progress
clawmonitor nudge --session-key 'agent:main:main' --template continue
clawmonitor push --session-key 'agent:main:main' --dry-run
clawmonitor status
clawmonitor status --detail
clawmonitor status --format json
clawmonitor status --format md
clawmonitor status --format md --detail
clawmonitor cron
clawmonitor models
clawmonitor models --mode direct --format json
clawmonitor models --mode openclaw --timeout 15
clawmonitor tree
clawmonitor report --session-key 'agent:main:main' --format both
clawmonitor watch --interval 1

Configuration

Default config path:

• ~/.config/clawmonitor/config.toml

Example config is in config.example.toml.

Optional labels

You can assign human-friendly names to long session keys (e.g. Feishu ou_...). See the [labels] section in config.example.toml.

Runtime data (NOT stored in this repo):

• Logs: ~/.local/state/clawmonitor/events.jsonl
• Reports: ~/.local/state/clawmonitor/reports/
• Cache: ~/.cache/clawmonitor/

Keys (TUI)

• ↑/↓: move selection
• PgUp/PgDn: page up / down
• g / G: jump to top / bottom
• Enter: nudge selected session (choose template)
• ?: show help overlay
• v: cycle sessions / models / system views
• s: jump directly to system view
• h: toggle status / history on the right
• u: cycle token windows (now / 1d / 7d / 30d)
• x: focus filter (hide stale sessions)
• t: toggle tree view (group by agent)
• c: toggle cron jobs in tree view
• R: rename/label selected session (writes [labels] in config)
• n: toggle NODE label mode (channel:label)
• l: toggle related logs panel
• d: re-run diagnosis for selected session
• e: export a redacted report for selected session
• z: cycle pane widths
• Z: toggle fullscreen detail pane
• o: open operator note in system view
• Esc: reset to the default surface
• r: force refresh
• f: cycle refresh interval
• q: quit

Rows are color-coded when your terminal supports colors (OK green, RUN cyan, IDLE yellow, ALERT red). In the details panel, Task: / Thinking: lines are highlighted (magenta when supported).

Model view notes:

• Model view is manual-refresh by design. Press r after switching with v.
• The top banner shows WAITING, RUNNING, DONE, or ERROR, so you can tell whether a probe run actually started.
• Each row shows the effective agent + model chain entry, including primary / fallbackN roles.
• ClawMonitor probes both paths when enabled:
  • Direct provider/API path (--mode direct or both)
  • OpenClaw execution path via temporary probe sessions (--mode openclaw or both)
• Supported direct transports today: openai-completions, openai-responses, anthropic-messages

Session/token/system notes:

• History loading is on demand. In the session view, press h and then r to read cached task history for the selected session.
• Token 1d / 7d / 30d windows are loaded from Gateway on demand and then cached in the TUI.
• System view is read-only by design: it summarizes service state, helper buildup, zombies/orphans, and reclaimable memory estimates without killing anything.

See docs/model-monitor.md for the probe model, classifications, and UI behavior. See docs/system-monitor.md and docs/system-view-guide.zh-CN.md for the new system view.

Telegram note: ACP “thread bindings”

OpenClaw can route a Telegram chat to a different session key via local thread bindings. This may make it look like your “main” session stopped receiving messages.

ClawMonitor detects this and flags it:

• BOUND_OTHER in clawmonitor status
• BIND in the TUI list

Relevant files/settings:

• Thread bindings: ~/.openclaw/telegram/thread-bindings-default.json
• Config toggle: ~/.openclaw/openclaw.json → channels.telegram.threadBindings.spawnAcpSessions

First run

If no config file exists, most commands will offer to run the init wizard (interactive terminals only).

See docs/launch-post.md for a longer intro.

Notes

• ClawMonitor never prints or writes OpenClaw secrets. It avoids dumping openclaw.json and redacts suspicious token-like strings in logs/reports.
• If Gateway is unreachable, ClawMonitor still works in offline mode (sessions/transcripts/locks/delivery-queue) but disables log tail + nudge.
• If your terminal window is narrow, clawmonitor tui may hide the details panel; use clawmonitor status as a stable fallback.
• For ClawHub import, see docs/clawhub-skill.md and skills/claw-monitor/SKILL.md.

See CONTRIBUTORS.md for acknowledgements.